Computing device and method

Abstract

The present disclosure provides a computation device. The computation device is configured to perform a machine learning computation, and includes an operation unit, a controller unit, and a conversion unit. The storage unit is configured to obtain input data and a computation instruction. The controller unit is configured to extract and parse the computation instruction from the storage unit to obtain one or more operation instructions, and to send the one or more operation instructions and the input data to the operation unit. The operation unit is configured to perform operations on the input data according to one or more operation instructions to obtain a computation result of the computation instruction. In the examples of the present disclosure, the input data involved in machine learning computations is represented by fixed-point data, thereby improving the processing speed and efficiency of training operations.

What is claimed is: 
     
       1. A computation device, comprising:
 a storage unit, 
 a controller unit, and 
 a conversion unit, wherein: 
 the controller unit is configured to:
 obtain a computation instruction; 
 parse the computation instruction to obtain one or more operation instructions, wherein each of the one or more operation instructions comprises an opcode field and an opcode, the opcode being configured to indicate information of a function of each operation instruction, and the opcode field comprising a first address and a length of first input data, a first address of output data, a decimal point position, a flag bit indicating a data type of the first input data, and an identifier for an operation type of the first input data; 
 for each of the one or more operation instructions, parse each of the one or more operation instructions to obtain the first address of the first input data, the first address of the output data, the decimal point position, the flag bit indicating the data type of the first input data, and the identifier for the operation type of the first input data; 
 obtain, for each of the one or more operation instructions, the first input data from the storage unit according to the first address of the first input data, and then transmit the first input data, the decimal point position, the flag bit indicating the data type of the first input data, and the identifier for the operation type of the first input data to the conversion unit; and 
 transmit the one or more operation instructions to an operation unit; and 
 
 the conversion unit is configured to convert, for each of the one or more operation instructions, the first input data into a second input data according to the decimal point position, wherein the flag bit indicates the data type of the first input data, and the identifier for the operation type of the first input data, wherein the second input data is fixed-point data; 
 wherein the computation device is configured to execute a machine learning computation, and further includes an operation unit, wherein:
 the conversion unit is further configured to transmit the second input data of each operation instruction to the operation unit, and 
 the operation unit is configured to operate the second input data of each operation instruction according to the one or more operation instructions to obtain a computation result of the computation instruction, and store the computation result into a storage space corresponding to the first address of the output data in the storage unit. 
 
 
     
     
       2. The computation device of  claim 1 , wherein the opcode field of each operation instruction includes the length of the first input data, and the controller unit is further configured to parse the operation instruction to obtain the length of the first input data, and
 wherein obtaining the first input data from the storage unit according to the first address of the first input data by the controller unit includes:
 obtaining, by the controller unit, the first input data from the storage unit according to the length of the first input data and the second input data. 
 
 
     
     
       3. The computation device of  claim 1 , wherein:
 the machine learning computation includes an artificial neural network operation, the first input data includes an input neuron and a weight, and the computation result is an output neuron. 
 
     
     
       4. The computation device of  claim 1 , wherein the operation unit includes a primary processing circuit and a plurality of secondary processing circuits, wherein:
 the primary processing circuit is configured to perform pre-processing on the second input data of each operation instruction and to transmit data and the one or more operation instructions between the plurality of secondary processing circuits and the primary processing circuit, 
 the plurality of secondary processing circuits is configured to perform an intermediate operation to obtain a plurality of intermediate results according to the second input data and the one or more operation instructions transmitted from the primary processing circuit, and to transmit the plurality of intermediate results to the primary processing circuit, and 
 the primary processing circuit is further configured to perform post-processing on the plurality of intermediate results to obtain the computation result of the computation instruction. 
 
     
     
       5. The computation device of  claim 4 , further comprising a direct memory access (DMA) unit, wherein:
 the storage unit includes any combination of a register and a cache, 
 the cache includes a scratch pad cache and is configured to store the first input data of each operation instruction, and 
 the register is configured to store scalar data in the first input data of each operation instruction, and 
 the DMA unit is configured to read data from the storage unit or store data in the storage unit. 
 
     
     
       6. The computation device of  claim 4 , wherein the controller unit includes an instruction cache unit, an instruction processing unit, and a storage queue unit, wherein:
 the instruction cache unit is configured to store the computation instruction associated with the artificial neural network operation, 
 the instruction processing unit is configured to parse the computation instruction to obtain the data conversion instruction and the one or more operation instructions, and to parse the data conversion instruction to obtain the opcode and the opcode field of the data conversion instruction, and 
 the storage queue unit is configured to store an instruction queue, the instruction queue including a plurality of operation instructions or computation instructions, wherein the plurality of operation instructions or computation instructions is to be executed in a sequence. 
 
     
     
       7. The computation device of  claim 6 , wherein the controller unit further includes:
 a dependency relationship processing unit configured to:
 determine whether there exists an associated relationship between a first operation instruction and a zeroth operation instruction before the first operation instruction, 
 cache the first operation instruction in the instruction cache unit based on a determination that there exists an associated relationship between the first operation instruction and the zeroth operation instruction, and 
 extract the first operation instruction from the instruction cache unit to the operation unit, when an execution of the zeroth operation instruction is completed, 
 
 wherein determining whether there exists an associated relationship between a first operation instruction and a zeroth operation instruction before the first operation instruction by the dependency relationship processing unit includes:
 extracting a first storage address interval of data required in the first operation instruction according to the first operation instruction, 
 extracting a zeroth storage address interval of data required in the zeroth operation instruction according to the zeroth operation instruction, 
 determining that there exists an associated relationship between the first operation instruction and the zeroth operation instruction, when an overlapped region exists between the first storage address interval and the zeroth storage address interval; and 
 determining that there does not exist an associated relationship between the first operation instruction and the zeroth operation instruction, when no overlapped region exists between the first storage address interval and the zeroth storage address interval. 
 
 
     
     
       8. The computation device of  claim 1 , wherein when the first input data is fixed-point data, the operation unit further includes:
 a derivation unit configured to derive a decimal point position of one or more intermediate results according to the decimal point position of the first input data of each operation instruction, wherein the one or more intermediate results are obtained by computing according to the first input data. 
 
     
     
       9. The computation device of  claim 8 , wherein the operation unit further includes:
 a data cache unit configured to cache the one or more intermediate results. 
 
     
     
       10. The computation device of  claim 1 , wherein the operation unit includes a tree module, and wherein:
 the tree module includes a root port coupled with the primary processing circuit and a plurality of branch ports coupled with the plurality of secondary processing circuits, and 
 the tree module is configured to forward data and the one or more operation instructions transmitted between the primary processing circuit and the plurality of secondary processing circuits, wherein the tree module is an n-tree structure, the n being an integer greater than or equal to two. 
 
     
     
       11. The computation device of  claim 1 , wherein the operation unit further includes a branch processing circuit, and wherein:
 the primary processing circuit is configured to:
 determine that the input neurons are broadcast data and the weights are distribution data, 
 divide the distribution data into a plurality of data blocks, and 
 transmit at least one of the plurality of data blocks, the broadcast data, and at least one of the one or more operation instructions to the branch processing circuits, 
 
 the branch processing circuit is configured to forward the data blocks, the broadcast data, and the plurality of operation instructions transmitted between the primary processing circuit and the plurality of secondary processing circuits, 
 the plurality of secondary processing circuits is configured to perform operations on the data blocks received and the broadcast data received according to the plurality of operation instructions to obtain a plurality of intermediate results, and to transmit the plurality of intermediate results to the plurality of branch processing circuits, and 
 the primary processing circuit is further configured to perform post-processing on the plurality of intermediate results received from the branch processing circuits to obtain a computation result of the computation instruction, and to send the computation result of the computation instruction to the controller unit. 
 
     
     
       12. The computation device of  claim 1 , wherein the plurality of secondary processing circuits is distributed in an array, wherein:
 each secondary processing circuit is coupled with adjacent other secondary processing circuits, and the primary processing circuit is coupled with K secondary processing circuits of the plurality of secondary processing circuits, 
 the K secondary processing circuits include n secondary processing circuits in the first row, n secondary processing circuits in the mth row, and m secondary processing circuits in the first column, and the K secondary processing circuits are configured to forward data and instructions transmitted between the primary processing circuit and the plurality of secondary processing circuits, 
 the primary processing circuit is further configured to: 
 determine that the input neurons are broadcast data, the weights are distribution data, 
 divide the distribution data into a plurality of data blocks, and 
 transmit at least one of the plurality of data blocks and at least one of the one or more operation instructions to the K secondary processing circuits, 
 the K secondary processing circuits are configured to convert the data transmitted between the primary processing circuit and the plurality of secondary processing circuits, 
 the plurality of secondary processing circuits is configured to perform operations on the data blocks according to the plurality of operation instructions to obtain a plurality of intermediate results, and to transmit the plurality of intermediate results to the K secondary processing circuits, and 
 the primary processing circuit is configured to process the plurality of intermediate results received from the K secondary processing circuits to obtain the computation result of the computation instruction, and to send the computation result of the computation instruction to the controller unit. 
 
     
     
       13. The computation device of  claim 10 , wherein:
 the primary processing circuit is configured to perform a combined ranking processing on the plurality of intermediate results received from the plurality of processing circuits to obtain a computation result of the computation instruction, or 
 the primary processing circuit is configured to a combined ranking processing and an activation processing on the plurality of intermediate results received from the plurality of processing circuits to obtain the computation result of the computation instruction. 
 
     
     
       14. The computation device of  claim 10 , wherein the primary processing circuit includes one or any combination of an activation processing circuit and an addition processing circuit, wherein:
 the activation processing circuit is configured to perform an activation operation on data in the primary processing circuit, and 
 the addition processing circuit is configured to perform an addition operation or an accumulation operation, 
 
       and wherein the plurality of secondary processing circuit includes:
 a multiplication processing circuit configured to perform a multiplication operation on the data blocks received to obtain a product result, and 
 an accumulation processing circuit configured to perform an accumulation operation on the product results to obtain the plurality of intermediate results. 
 
     
     
       15. A machine learning operation device, comprising one or more computation devices according to  claim 1 , wherein the at least one computation device is configured to obtain data to be processed and control information from other processing devices, to perform specified machine learning computations, and to transmit an execution result to the other processing devices through I/O interfaces, wherein:
 when the machine learning operation device includes a plurality of the computation devices, the plurality of computation devices is configured to couple and transmit data with each other through a specific structure, and the plurality of computation devices is configured to:
 interconnect and transmit data through a fast external device interconnection PCIE (peripheral component interface express) bus to support larger-scale machine learning computations, 
 share the same one control system or have respective control systems, 
 share the same one memory or have respective memories, and 
 deploy an interconnection manner of any arbitrary interconnection topology. 
 
 
     
     
       16. A computation method implemented in a computing device according to  claim 1 , comprising:
 obtaining, by the controller unit, the one or more operation instructions; 
 parsing, by the controller unit, each operation instruction to obtain, for each operation instruction, the first address of the first input data, the first address of the output data, the decimal point position, wherein the flag bit indicates the data type of the first input data, and the identifier for the operation type of the first input data; 
 obtaining, by the controller unit, for each operation instruction, the first input data from the storage unit according to the first address of the first input data, and 
 converting, by the conversion unit, for each operation instruction, the first input data into a second input data according to the decimal point position, wherein the flag bit indicates the data type of the first input data, and the identifier for the operation type of the first input data, wherein the second input data is fixed-point data. 
 
     
     
       17. The method of  claim 16 , wherein obtaining the one or more operation instructions by the controller unit includes:
 obtaining, by the controller unit, the computation instruction, and parsing, by the controller unit, the computation instruction to obtain the one or more operation instructions. 
 
     
     
       18. The computation device of  claim 17 , wherein the opcode field of each operation instruction further includes the length of the first input data, and
 wherein obtaining the first input data according to the first address of the first input data by the controller unit includes obtaining the first input data according to the first address and the length of the first input data. 
 
     
     
       19. The method of  claim 18 , wherein the method is configured to execute the machine learning computation, and further includes:
 operating, by the operation unit, the second input data of each operation instruction according to the one or more operation instructions to obtain the computation result of the computation instruction, and 
 storing, by the operation unit, the computation result into the storage space corresponding to an output address of the data. 
 
     
     
       20. The method of  claim 18 , wherein the machine learning computation includes the artificial neural network operation, the first input data of each operation instruction includes an input neuron and a weight, and the computation result is an output neuron. 
     
     
       21. The method of  claim 18 , wherein, for each operation instruction, when the first input data and the second input data are both fixed-point data, the decimal point position of the first input data is inconsistent with that of the second input data. 
     
     
       22. The method of  claim 21 , wherein, for each operation instruction, when the first input data is fixed-point data, the method further includes:
 deriving, by the operation unit, a decimal point position of one or more intermediate results according to the decimal point position of the first input data, wherein the one or more intermediate results are obtained by operating according to the first input data.